The Agent Skills Directory

[COMMAND_EXECUTION]: The skill triggers the execution of local shell scripts, specifically tunnel-runtime.zsh and mcp-workspace-launcher.zsh, to manage a secure tunnel and validate workspace access. These scripts are executed from a user-defined path specified in the local configuration.
[DATA_EXFILTRATION]: The orchestration involves transmitting project code, local diffs, and test results to a third-party service (ChatGPT) for review. This data transmission is the primary function of the skill and is constrained to a read-only environment, with explicit instructions to redact secrets and private identifiers.
[INDIRECT_PROMPT_INJECTION]: The skill operates on data ingested from external web searches and model outputs, which could potentially contain adversarial instructions aimed at influencing local implementation tasks.
Ingestion points: External web content retrieved via search and assistant-generated architectural plans.
Boundary markers: The skill uses a structured communication protocol with unique task IDs and specific reply markers to validate response boundaries and prevent spoofing.
Capability inventory: The agent has the authority to execute local scripts, modify codebases, and interact with the web via UI automation.
Sanitization: The instructions mandate the sanitization of all artifacts, redaction of sensitive URLs, and verification of all external claims against the local environment prior to implementation.

yichen-codex-chatgpt